A folder of screenshots won't change what you make. A register of claims, proof and offers will, and AI can build it without you copying a thing.
Public ad libraries show what competitors are running, not what is working. Capture ads in small, dated batches, have AI fill a seven-column register from the text, and ask it for patterns and gaps. Build your angle from proof you can show, and score every draft with the six-question copy check. Libraries have their own terms, and this is not legal advice.
Competitor ad research tends to end in a folder. Screenshots go in, a few get passed around with a note saying we should do something like this, and a month later the team’s own ads look exactly as they did. A screenshot records what an ad looks like. It says very little about what the ad is claiming, who it’s talking to, what proof it offers and what it asks the viewer to do, and those four things are the part you can learn from.
AI is useful in the gap between the screenshot and the decision. Give it the text of a stack of ads and it will sort them by claim, proof and offer faster than a person with a spreadsheet, and notice that several competitors lead with speed while none mentions support. It can’t tell you which ads are working or choose your angle, and it shouldn’t write something like theirs. This guide covers how to analyze competitor ads with AI on those terms: a method, prompts to paste, a register to keep, and a line between learning and copying.
The raw material is public. Meta’s help page says its Ad Library contains all active ads shown across Meta technologies, and that anyone can view and search it.[1] Google’s Ads Transparency Center was launched in 2023 as, in Google’s words, a searchable hub of ads served from verified advertisers, where you can see the ads an advertiser has run, the region they were shown in, the last date an ad ran and its format.[2] A June 2023 Google policy update extended this to all advertisers serving ads on Google platforms, including ones not yet verified.[3] Both change, so the help pages are the authority on what’s there today. What matters is what each leaves out.
| Library | What you can see | What the page doesn’t give you |
|---|---|---|
| Meta Ad Library | The creative and copy of active ads, the Page that ran them, the date an ad started running, and Page history such as name changes | For ordinary commercial ads outside the EU, the UK and the political categories: no spend, impressions, reach or targeting. Ads delivered in the EU also show total reach and ad-level audience selections |
| Google Ads Transparency Center | The ads an advertiser has run, the region they were shown in, the last date an ad ran and the ad’s format | Google’s launch post doesn’t list spend, results or targeting. The 2023 policy page says some extra detail, such as targeting, may appear where the law requires it |
From Meta’s help page and Google’s launch post and policy update, read October 15, 2026.
A paid social specialist opens a competitor’s Page and finds one headline running in many variants, with the oldest still live. It’s tempting to read that as a winner. What the library supports is narrower: the advertiser has kept paying to show it, and nothing on the page says what it returned.
A live ad shows what a competitor is paying to show, not what is paying them back.
The fix for a folder is a register: one row per ad, the same columns for each ad, filled from what is visibly in the ad, with one separate column for your own read. It turns a pile of ads into something you can compare, and it gives AI a structure to fill instead of a mood to summarise.
| Ad | Claim | Audience cue | Proof | Offer and ask | Hook and format | Our read (inference) |
|---|---|---|---|---|---|---|
| A-1, captured week 1 | Cuts scheduling time | Office managers (“your front desk”) | Customer logo strip, no result shown | Trial, “Start now” | Static image, headline on a plain background | Leads with time saved. Borrows credibility from logos rather than results |
| B-1, captured week 1 | Fewer no-shows | Clinic owners | A named customer quote | “Book a demo” | Short video, problem shown first | Specific audience and pain. Asks for a bigger commitment |
| C-1, captured week 1 | Switch without losing your data | People already using a rival tool | A migration promise, nothing shown | Import your calendar, “Switch today” | Carousel, comparison framing | Aimed at switchers. The claim is untested inside the ad itself |
Illustrative rows for a made-up scheduling app category, not real ads. Only the last column is allowed to guess.
The two columns people tend to skip are Proof and the capture date. Without proof you end up comparing promises, and promises all sound alike. Without the date you can’t tell an ad that’s been live for months from one that appeared last week.
The last column is what makes AI safe to use here. If the model has to put every interpretation in its own labelled place, an inference can’t quietly pass as a fact. Run the prompt below on each competitor’s ad text, and keep the output only after checking it against the originals.
| I’m building a competitor ad register. Below is the text of [NUMBER] ads from [COMPETITOR], copied on [DATE] from the [Meta Ad Library / Google Ads Transparency Center]. For each ad, fill these fields using only the words I pasted: Claim (the promise, in under 12 words); Audience cue (who the copy addresses, quoting the words that show it); Proof (what supports the claim, or “none shown”); Offer and ask; Hook (the first line). If a field isn’t in the text, write “not stated”. Do not guess performance, spend, targeting or results, and do not describe any image I haven’t described. Put your interpretation in a separate column called Our read, one sentence per ad, and start each with “Inference:”. [PASTE AD TEXT, ONE AD PER BLOCK, LABELLED A-1, A-2 …] |
Check three rows against the original ads before you trust the rest. If a field is wrong in one, assume it’s wrong in more.
For video and image ads, write the hook and any on-screen text into the paste yourself. If your tool accepts images, check what it says it read, because a misread line of overlay text flows straight into your claim column.
Collecting is the dull part, and where trouble starts if you automate it. I’m not walking through the screens, because they change and each library’s help page explains its current steps better than a blog post can. The habits matter more than the clicks.
Meta’s Terms of Service say you may not access or collect data from its products using automated means without its prior permission.[4] Its help page says you can download Ad Library searches, but only when logged in to Facebook and having agreed to its terms, and that the Ad Library API is aimed at ads about social issues, elections or politics and ads delivered to the EU and associated territories.[1] For ordinary commercial ads, that points to manual capture in small batches. Check any library’s current terms before pointing a scraper or browser extension at it, and ask counsel if the plan is volume.
A content marketer returns to a rival’s Page after a few weeks and finds that a good share of the ads she saw last time are gone. The register from the first visit is now the only record of them, which is the case for dating every capture.
That changes what the register is for. A one-off teardown shows what a competitor says today. Dated captures start to show what they keep, change and stop saying, which is about as close to a read on their strategy as public information allows. It stays a read, not a result.
With a few weeks of rows, the job shifts from collecting to comparing. Comparing dozens of rows for shared claims and missing proof is slow by hand and quick for a model, as long as it’s fenced in. The prompt below keeps it to the rows and stops it inventing performance.
| Here is my competitor ad register for [CATEGORY]: [PASTE THE FULL REGISTER]. Work only from these rows. (1) Group the claims into no more than five themes and list which rows sit in each. (2) For each theme, say what proof is offered and where the proof is “none shown”. (3) List the audiences the copy addresses, and any audience that no row addresses. (4) Separate what the rows show from what you are inferring, and mark every inference. (5) List three things this register cannot tell us, such as performance, spend or targeting. Do not say or imply that an ad performs well because it has been running. Do not invent ads that aren’t in the register. If two rows contradict each other, say so rather than smoothing it over. |
The “cannot tell us” list is the useful part. It reminds the whole team what the register is and isn’t.
Read the answer for three things. Can every theme be traced to specific rows? A theme with no rows behind it is the model filling space. Did the proof column survive? Models like to call a claim “well supported” when the row says none shown. And has targeting crept in? The libraries mostly don’t show it for ordinary commercial ads, so “they’re clearly targeting small clinics” is a guess built from copy. If a confident claim looks off, the cause is usually the one covered in why AI hallucinates: fluent text fills a gap.
For sizing up rivals beyond their ads, see our competitor research playbook and step-by-step competitor analysis guide. This piece stays with the ads.
A pattern summary is not a plan. The step that earns the work is turning it into gaps, of which there are three kinds:
A brand manager reviewing the summary often finds the crowded theme is the one the team’s own ads lead with. The next prompt turns a gap into angles you could test, built only from what you can prove. Expect it to hand back some angles you can’t support yet, and treat those as a to-do list for evidence, not as copy to write.
| We sell [WHAT WE SELL] to [AUDIENCE]. The only proof we can show is: [OUR REAL PROOF, e.g. a guarantee, a documented result, a policy, a customer quote we have permission to use]. Here is the pattern summary of our competitors’ ads: [PASTE SUMMARY]. Suggest four angles we could test. For each give the claim, the audience, the proof from my list only, and how it differs from the competitor themes. Flag any angle that would need proof we don’t have, and say what that proof would be. Rules: do not reuse any phrase of five or more words from the competitor ads; do not imitate their layout, imagery or script; do not mention competitor names in the copy; do not make a claim our proof doesn’t support. |
The proof line is what keeps the angles honest. If it’s empty, the output is a list of things to go and find out.
Pick one or two angles, not four. Writing the ad is a separate job, covered in writing Facebook ad copy with AI and writing Google Ads with AI. Whether the angle works is a question for your own test, and our guide to testing ad creative covers the tools.
Learn the pattern across ads, and leave the asset behind.
The line between learning and copying is partly law, partly judgment, and the law differs by country. I’m not a lawyer, and this is not legal advice. What I can do is lay out what the primary sources say and where I’d stop and ask someone who is. Past the law, I’d use a plain test: if the rival’s marketing lead read your ad next to theirs, would they see a competitor who did their own thinking, or someone who borrowed theirs?
Start with copyright. US law says protection for an original work doesn’t extend to any idea, procedure, process, system, method of operation or concept, regardless of how it’s described or illustrated.[5] The Copyright Office adds that names, titles, slogans and short phrases aren’t protected by copyright, though a logo’s artwork may be, and names and slogans may be trademarks instead.[6] In practice that means a structure such as problem, proof, offer is fair to learn from. A competitor’s photograph, video script, layout or distinctive wording is a different matter, and an “inspired by” version produced by AI is still a copy with an extra step.
Trademark is the second risk. The USPTO describes the core problem as consumers being likely to mistakenly believe goods or services come from the same source, and notes that marks don’t have to be identical to be confusingly similar.[7] A lookalike name, colour scheme or logo treatment, or copy that suggests a connection with a rival, is the version to avoid. Comparative ads that name a competitor have country-specific rules, so those go to counsel.
The third risk is the least obvious: borrowing a claim. The FTC says the law requires advertisers to have proof before the ad runs, and defines a reasonable basis as objective evidence that supports the claim.[8] A rival’s claim may rest on evidence you don’t have, so copying “fastest setup” moves their claim, and the burden of proving it, onto you.
Borrow the question an ad answers, not the claim it makes.
| Fair to learn from | Where the risk starts |
|---|---|
| The structure: problem, proof, offer | Their photograph, video script or layout |
| A theme several competitors share | Their distinctive wording, or an AI-made “inspired by” version of it |
| The question an ad answers | Their claim, without your own evidence |
| What category buyers respond to | Names, colours or logo treatment that suggest a connection |
A Future Factors working view of the line, not legal advice.
Before a draft goes anywhere, score it. Each line is 0 (no), 1 (partly) or 2 (yes).
| # | Question | What a 2 looks like |
|---|---|---|
| 1 | Does the idea show up across several advertisers? | You can point to at least three competitors’ ads, so it’s a pattern and not one ad’s asset |
| 2 | Is the wording your own? | No run of five or more words matches a competitor’s ad |
| 3 | Does it look different? | Flip between the two and layout, image and script are clearly not the same |
| 4 | Could a customer confuse the two businesses? | Name, colours, logo treatment and tone give no hint of a connection |
| 5 | Is every claim backed by your evidence? | You can show proof for each claim you make, and it isn’t theirs |
| 6 | Could you explain the angle without showing their ad? | The idea stands on its own, in your words |
Maximum 12. Thresholds are a Future Factors working guide, not a legal standard. 10 to 12: go. 7 to 9: rework the lowest line. 0 to 6: stop, and take anything scoring low on lines 3 or 4 to counsel.
A high score doesn’t clear you legally. The check is a prompt for judgment, not a substitute for advice, and how you collect the ads belongs in the same conversation as what you publish.
The register only helps if it’s clear who is responsible for what. AI does the sorting, you own what the pattern means, and someone checks the output against the originals before it reaches a brief.
| What AI does | What the human still owns | How it gets checked |
|---|---|---|
| Fills the register from pasted ad text and flags where a field is not stated | Choosing which competitors and which ads to capture, and writing the hook for video ads | Three rows checked against the original ads each time the register is updated |
| Groups claims into themes and lists proof gaps and missing audiences | Deciding which gaps matter for your business and which to leave alone | Every theme traced back to named rows; any theme without rows is removed |
| Drafts angles from the proof you supply | Deciding what you can honestly claim, and choosing the one angle to test | The copy check scored by someone other than the drafter, with a named person who signs off |
A Future Factors working view.
The emphasis shifts by seat.
| Role | What to focus on | Where to be careful |
|---|---|---|
| Paid social or search specialist | Hooks, formats and how offers are framed, captured on a regular schedule | Reading long-running ads as winners without your own test |
| Brand manager | Claims, tone and what the category all sounds like | Drifting toward a rival’s look without noticing |
| Content marketer or copywriter | Proof and audience gaps that become angles | Reusing register wording in drafts |
| Founder or small-business owner | Three rivals, one gap, one test | Skipping the copy check with no one to ask |
A Future Factors working view.
Counting rows added is the weakest signal that the habit is paying off.
| Question | What to look at |
|---|---|
| Use | Is the register updated on schedule? |
| Persistence | Is it still updated two months later, without a reminder? |
| Impact | Did it change a brief, a test, or something you decided not to run? |
Impact justifies the time.
Teams can set this up on their own: one person, a shared sheet, a monthly slot. Training with an experienced facilitator shortens the learning curve, because spotting a weak inference and holding the line on the copy check are skills a team picks up faster together. In our terms, training builds the capability, workflow design gives it a home in the dated register and review slot, and reinforcement, like a manager asking what changed because of it, turns it into a habit.
The output is a short note, not a deck: what changed, what you infer (labelled), the one angle you’ll test, and the one thing you’re deliberately not copying, so the reason a tempting offer was left alone is on record.
| Field | Example |
|---|---|
| What changed | Two of the three rivals added ads that lead with migration help. One dropped its logo-strip ad. |
| What we infer | Inference: the category is chasing switchers. We can’t tell whether it’s working for them. |
| What we’ll test | One angle built on our documented onboarding timeline, in our own words. |
| What we’re not doing | Matching their offer wording or look. The copy check scored low on lines 2 and 3 for that draft. |
| Owner and review date | A named person, at the next monthly slot. |
Illustrative content.
Five steps for week one.
Put a repeat date in the calendar. The second teardown is where patterns begin to show.
Looking at ads a platform publishes in its own public library is ordinary market research. The risk sits in what you publish afterwards (copied wording, layouts, lookalike branding, claims you can’t prove) and in how you collect. This is not legal advice, and rules differ by country, so ask counsel before anything borderline goes live.
Meta’s Terms of Service say you may not access or collect data from its products by automated means without its prior permission, and its Ad Library API focuses on political and social-issue ads and ads delivered to the EU. For ordinary commercial ads, capture manually in small batches and read each library’s current terms first.
There’s no threshold I can source, and the libraries don’t show results. A long-running ad tells you the advertiser kept it live. Treat it as a reason to look closer, then test your own idea.
It depends on the tool, and I haven’t tested every one, so check what yours accepts and what it says it read. For the register, paste the ad text and write the hook and on-screen text of video ads yourself, because misread overlay text becomes a wrong claim.
Run the copy check. If wording or look is close, rewrite from your own proof and change the claim and the visual, not just the phrasing. If the closeness involves a name, logo or colour treatment, take it to counsel before it runs.
Sources were read on October 15, 2026: Meta’s Business Help Center page About the Meta Ad Library (including its table of what is shown for different kinds of ads), Meta’s Terms of Service, Google’s March 2023 announcement of the Ads Transparency Center, Google’s Advertising Policies Help page on the June 2023 ads transparency policy update, the Copyright Office FAQ on what copyright protects, 17 U.S.C. 102, the USPTO page on likelihood of confusion and the FTC’s Advertising FAQs for small businesses. I did not describe any library’s screens or filters, because they change and I did not verify them. The teardown register, the copy check, its score bands and the role table are Future Factors working methods based on our own judgment, not research-validated models. The example rows are illustrations, not real ads or client stories. We did not use any figure for how long successful ads run or what competitor ads cost, because we could not confirm one against a primary source. This is not legal advice.
Weekly tips. Real examples. Practical help for busy professionals.
We care about your data, check out our privacy policy.